ieee communications magazine, in press, 2017 · pdf fileieee communications magazine, in...
TRANSCRIPT
IEEE COMMUNICATIONS MAGAZINE, IN PRESS, 2017
Wireless-Optical Network Convergence: Enabling the 5G Architecture
to Support Operational and End-User Services
Anna Tzanakaki(1),(12),Markos Anastasopoulos(1), Ignacio Berberana(2), Dimitris Syrivelis(3), Paris
Flegkas(3),Thanasis Korakis(3), Daniel Camps Mur (4), Ilker Demirkol(5), Jesús Gutiérrez(6), Eckhard Grass(6),
Qing Wei (7), Emmanouil Pateromichelakis(7), Nikola Vucic(7), Albrecht Fehske (8), Michael Grieger (8),
Michael Eiselt(9), Jens Bartelt(10), Gerhard Fettweis(10), George Lyberopoulos(11), Eleni Theodoropoulou(11)
and Dimitra Simeonidou(1)
(1) HPN Group, University of Bristol, UK, (e-mail: [email protected]), (2)Telefonica Investigacion Y Desarrollo, Madrid, ESP, (3) University of Thessaly, Volos, GRC, (4) i2CAT Foundation, ESP, (5) Univeristat Politecnica de Catalunya, ESP, (6)IHP GmbH, DE(7) Huawei Technologies Duesseldorf GmbH, DE, (8)Airrays GmbH, Dresden, DE, (9) ADVA Optical Networking SE, (10) Vodafone Chair Mobile Communications Systems, Technische Universität Dresden, DE, (11) COSMOTE Mobile Communications S.A, R & D Dept., Fixed & Mobile, GRC, (12) National and Kapodistrian University of Athens, Department of Physics, GRC
Abstract— This paper presents a converged 5G network infrastructure and an overarching
architecture, to jointly support operational network and end-user services, proposed by the EU 5G PPP
project 5G-XHaul. The 5G-XHaul infrastructure adopts a common fronthaul/backhaul network
solution, deploying a wealth of wireless technologies and a hybrid active/passive optical transport,
supporting flexible fronthaul split options. Τhis infrastructure is evaluated through a novel modeling.
Numerical results indicate significant energy savings at the expense of increased end-user service delay.
Keywords—5G, backhauling, fronthauling, small cells, C-RAN
I. INTRODUCTION
The enormous growth of mobile data predicted is attributed to the rapidly increasing: a) number of network-
connected end devices, b) Internet users with heavy usage patterns, c) broadband access speed, and d)
popularity of applications such as cloud computing, video, gaming etc. Traditional Radio Access Networks
(RANs), where Base Band Units (BBUs) and radio units (RUs) are co-located, cannot meet this massive
foreseen growth. This is attributed to high capital and operational costs associated with the lack of resource
sharing and modularity, reduced agility and scalability as well as inefficient energy management.
Cloud Radio Access Networks (C-RANs) propose to overcome these limitations, by supporting connection
of Access Points (APs), known as RUs, to a BBU pool hosted in a Central Unit (CU) through a set of transport
IEEE COMMUNICATIONS MAGAZINE, IN PRESS, 2017
links. These links are referred to as fronthaul (FH). Currently interfacing between RUs and CU is enabled
through the adoption of standards such as the Common Public Radio Interface (CPRI). The RU wireless signals
are commonly transported over an optical FH network, using either digital transmission (e.g. CPRI), or analog
transmission (radio-over-fiber). The adoption of CPRI type of solutions enables consolidation of a larger
number of BBUs per CU by extending the transport network range. However, C-RAN requires very high
transport bandwidth due to the traffic volume created by the sampled radio signals transported to the CU and
the very tight delay and synchronization specifications [1]. Existing mmWave E-Band and optical transport
solutions supporting traditional backhaul (BH) requirements are based on different flavours of Passive Optical
Networks (PONs) and 10GE technologies. Considering that in 5G environments these transport solutions will
also need to offer FH capabilities, it is clear that they will be unable to offer the required capacity for both BH
and FH services. To take advantage of the benefits and address the challenges associated with C-RAN,
equipment vendors are expanding their FH solutions adopting advanced wireless technologies (e.g. Sub-6GHz
and 60GHz bands, including advanced beam tracking and MIMO techniques), and new flexible and dynamic
Wavelength Division Multiplexing (WDM) optical networks [2]. These are also enhanced with novel control
and management approaches to enable increased granularity, end-to-end optimization and guaranteed Quality
of Service (QoS).
To facilitate CRAN’s technical feasibility and benefit from its coordination and pooling gains there is a need
to relax the FH requirements. In view of this, solutions proposing FH compression and alternative architectures
relying on flexible functional splits (Figure 2) have been reported [3], [4]. The concept of flexible splits relies
on transferring some of the processing functions away from the RU and locating these centrally at a CU. These
functions are commonly performed through dedicated and specific purpose hardware, with significant
installation, operational and administrative costs. To address these issues, the concept of network softwarisation
enabling migration from traditional closed networking models to an open reference platform able to instantiate
a variety of network functions has been recently proposed. A typical example includes the OpenAirInterface
(OAI) i.e. an open source 4G/5G radio stack able to be executed on general purpose servers hosted in data
centers (DCs) [5]. Such open source frameworks are still in early development stages and do not allow
execution of more complex functionalities such as flexible RAN splits. In this study, the concept of flexible
functional splits is addressed by appropriately combining servers with low processing power (cloudlets) and
relatively large-scale DCs placed in the access and metro domains respectively. The remote processing
requirements associated with some of the functional split options, impose the need for a high bandwidth
transport interconnecting RUs and the CU. On the other hand, the variability of remote processing requirements
IEEE COMMUNICATIONS MAGAZINE, IN PRESS, 2017
across the various split options introduce the need for a transport network that offers finely granular and elastic
resource allocation capabilities.
Figure 1:The 5G-XHaul Physical Infrastructure: FH and BH services are provided over a common wired/wireless network
infrastructure. In the FH case, parts of the BBU processing can be performed locally and some parts remotely at the DCs enabling
the C-RAN flexible split paradigm. BBUs are executed in general purpose servers in the form of virtual entities. BH services
interconnect end-users with Virtual Machines hosted in the DCs.
Addressing these challenges, we propose a network solution that converges heterogeneous network domains
deploying optical and wireless technologies together with compute resources in a common 5G infrastructure.
This infrastructure, developed in the framework of the EU 5G PPP project 5G-XHaul, will support both
operational network as well as fixed and mobile end-user services. Operational network services refer to
services required for the operation of the 5G infrastructure e.g. FH services offered to infrastructure
operators/providers. On the other hand, end-user services refer to services provided to end users (e.g. content
delivery, gaming etc) that in 5G environments require BH connectivity, referred to as BH services. The main
technical innovations of the proposed solution include: i) an architectural framework aligned with the Software
Defined Networking (SDN) open reference architecture [6] and the ETSI Network Function Virtualization
(NFV) standard to jointly support FH and BH services as well as the adoption of flexible functional split
options. This is a key innovation of the proposed architecture compared to contemporary LTE-A systems where
FH and BH services are supported by separate and dedicated networks, while network control and management
is closed, ii) a novel data plane design that converges heterogeneous wireless and optical solutions and, iii) a
VM
vBBU2
DataCenters
SmallCells
RU
RU
RU
RU
eNB
RU
RURURU
EPCSGW/GSN
vBBU
vBBUvBBU
PDN-GW
vBBU1
Backhaul
Fronthaul
vBBU
vBBU1 vBBU2
VM
eNB
MacroCell
Internet
60GHz/Sub-6linksforFH/BHFHBH
HeNBFemtoCells
HeNBHeNBGW
HeNB
WiFiWLANGW
TSON
WDM-PON
5G-XhaulWirelessBackhaul/Fronthaul
HeNB:HomeeNodeB
WirelessAccess
VM:VirtualMachineGW:Gateway
vBBU:VirtualBaseBandUnitRU:RemoteUnitPDN-GW:PacketDataNetworkGateWay
EPC:EvolvedPacketCoreSGW:ServingGateWay
IEEE COMMUNICATIONS MAGAZINE, IN PRESS, 2017
novel modelling framework adopting multi-objective optimization (MOP) techniques to evaluate the proposed
architectural approach. This modelling framework focuses on optimal FH and BH service provisioning, with
the overall objective to maximize the infrastructure energy efficiency and, minimize end-to-end service delays.
II. OVERVIEW OF THE 5G-XHAUL ARCHITECTURE
A. Data Plane Architecture
The 5G-XHaul data-plane design considers converged optical and wireless network domains in a common
5G infrastructure supporting both transport and access. In the wireless domain, a dense layer of small cells can
be wirelessly backhauled through mm-Wave and sub-6 technologies. Alternatively, small cells can be
connected to a CU through the 5G-Xhaul optical network. This adopts a hybrid approach combining a dynamic
and elastic frame based optical network solution with enhanced capacity WDM PONs [7], to support the
increased transport requirements of 5G environments in terms of granularity, capacity and flexibility.
Given that 5G-XHaul proposes the adoption of C-RAN to overcome traditional RAN limitations (Figure 1),
it introduces the need to support new operational network services (FH) over the transport network. These
emerge from the need to connect densely distributed RUs with the CU, meeting very tight latency and
synchronization requirements. To maximize coordination and resource sharing gains 5G-XHaul proposes to
support BH and FH jointly in a common infrastructure. Thus achieving improved efficiency and management
simplification leading to measurable benefits in terms of cost, scalability and sustainability. Aiming to address
the C-RAN challenges described above, 5G-XHaul proposes the adoption of flexible split options that can relax
the tight transport requirements in terms of capacity, delay and synchronization. Figure 2 shows the range of
“optimal split” options, that span between the “traditional distributed RAN” case, where “all processing is
performed locally at the AP”, to the “fully-centralized C-RAN” case, where “all processing is allocated at the
CU”. All other options allow allocating some processing functions at the RU, while the remaining processing
functions are performed remotely at the CU. The optimal allocation of processing functions to be executed
locally or remotely i.e. the optimal “split”, can be decided based on factors such as transport network
characteristics, network topology and scale as well as type and volume of services.
A key enabler of the 5G-XHaul data plane (Figure 1) is the hybrid (passive-active) optical network transport
that supports jointly FH and BH services offering the required connectivity, capacity and flexibility. The
passive solution employs WDM-PONs, while the active solution adopts the highly versatile Time-Shared
Optical Network (TSON) [7] extended to support novel features offering fine bandwidth granularity (variable
length optical frames) and elastic bandwidth allocation capabilities.
IEEE COMMUNICATIONS MAGAZINE, IN PRESS, 2017
Given the technology heterogeneity supported by the 5G-XHaul data plane, a critical function of the
converged infrastructure is interfacing between technology domains. Interfaces are responsible for handling
protocol adaptation as well as mapping and aggregation/de-aggregation of traffic across domains. Different
domains (wireless/optical) may adopt different protocol implementations and provide very diverse levels of
capacity (Mbps for the wireless domain to tens of Gbps for TSON), granularity (Kbps for the wireless domain
to 100 Mbps for TSON) etc. A key challenge also addressed by these interfaces is mapping of different QoS
classes across different domains as well as flexible scheduling enabling QoS differentiation mechanisms. More
specifically, at the optical network ingress point (e.g. TSON edge node) the interfaces receive traffic frames
generated by fixed and mobile users and arrange them to different buffers. The incoming traffic is aggregated
into optical frames and is assigned to suitable time-slots and wavelengths according to the adopted queuing
policy, before transmission in the TSON domain. For FH traffic a modified version of the CPRI protocol
supporting the concept of functional splits (eCPRI) has been adopted. Note that due to the large variety of
technologies involved in 5G, these interfaces need to support a wide range of protocols and technology
solutions and execute traffic forwarding decisions at wire-speed. This requires the development of
programmable network interfaces combining hardware level performance with software flexibility. The reverse
function is performed at the egress TSON edge node. More information regarding interfacing of wireless and
optical domains is available at [7], [13].
B. Overarching Layered Architecture
Managing and operating heterogeneous infrastructures integrating a variety of optical and wireless
technologies and domains such as the 5G-XHaul infrastructure presents several challenges. To address these
we propose the adoption of the integrated SDN/NFV paradigm. This will take advantage of the separation of
control and data plane offered by SDN and the depoyments of the variety of NFV elements. Through this
integration the benefits of the control and the holistic network view of SDN will be combined with the
flexibility to provision services by composing Service Chains (SCs) through orchestrated network functions.
SDN/NFV integration allows SDN controllers to control the Virtual Network Functions (VNFs) [11] enabling
on-demand resource allocation for dynamically changing workloads [6]. SDN network elements may
correspond to both Physical Network Functions (PNFs) and VNFs if they are implemented in virtualized
environments, as software running on general-purpose hardware platforms [6]. The virtualization of network
elements enables flexible allocation of data plane resources according to network applications requirements.
On the other hand, SCs offering orchestrated service provisioning over heterogeneous environments are
considered to be a possible network application, which can include SDN controller functions or interact with
SDN controllers to provide VNFs.
IEEE COMMUNICATIONS MAGAZINE, IN PRESS, 2017
Figure 2: The overall overarching architecture supporting functional split processing [3]-[4].
The details of the 5G-XHaul overarching architecture adopting the integrated SDN/NFV paradigm to
facilitate management and operation of the heterogeneous physical infrastructure (PI) are illustrated in Figure
2.
The Infrastructure Management Layer (IML) manages the different technology domains. This layer is
responsible to enable multi-tenant operation through cross-domain slicing and virtualization facilitating joint
FH and BH services over the common infrastructure. Information retrieval and communication between
domains is handled by network and compute controllers located at this layer, enabling resource abstraction and
virtualization. Therefore, IML supports traditional management of the PI together with advanced features
required for virtualization and virtual resource management functions.
The Control Layer (CL) is responsible for cross-domain orchestration of virtual and PIs, created and exposed
by the IML having an overall view of all network domains. CL provides end-to-end connectivity services in
the form of SCs deploying converged control and management procedures with guaranteed QoS. CL supports
configuration of both virtualized and non-virtualized heterogeneous resources as well as legacy devices,
EndPointA
VM
EndPointB
vBBU(upperlayers)
VirtualWireless
forwarding
VirtualOptical
VirtualBH
VirtualWireless
forwarding
VirtualOptical
VirtualFH
RRHRRH
RRH
VM
vBBUInternet
Fronthaul
OpticalTransportDataCenters
Backhaul
DataCenters
vBBU
WirelessAccess/Transport
VM
mmWaveNetworkController
LTENetworkController
ComputeController
WDMPONController
TSONController
ComputeController
Virtualization Virtualization Virtualization Virtualization
VirtualNetwork
PhysicalNetwork
Processing VirtualNetwork
PhysicalNetwork
Processing
SDNController VNF SDN
Controller VNF
EM EMElementManagement
(EM) ElementManagement(EM)
DriversMan
aged
PHY
Infrastructure
Infra
struc
ture
Man
agem
ent
RU
RFtoBaseband
CyclePrefix&FFT
Resourcedemapping
Receiveprocessing
Decoding
MAC
(1)
(2)
(3)
(4)
(5)
Clou
dRA
NTrad
ition
alR
AN• highnetworkbandwidth• IncreasedBBUsharing
• lownetworkbandwidth• LimitedBBUsharing
Control
Man
agem
ent&
Service
Orche
stratio
n
vBBU(lowerlayers)
IEEE COMMUNICATIONS MAGAZINE, IN PRESS, 2017
through a set of distributed SDN controllers and facilitates the development of enhanced VNFs to operate the
5G infrastructure seamlessly.
The Management and Service Orchestration Layer (MSOL) handles orchestration requirements for the
delivery of network and compute services as well as composition and provisioning of SCs in multi-tenant
environments deploying VNFs. MSOL is also responsible to support interoperability with legacy software and
hardware.
III. USE CASE: JOINT OPTIMIZATION OF FH/BH
As already discussed the 5G-XHaul data plane architecture can jointly support FH and BH services adopting
a hybrid optical transport, integrating passive and active optical networks. The higher layers of the architecture
that facilitate access and management of both network and compute resources also play a key role. The ability
of the IML to create virtual infrastructure (VI) slices across heterogeneous domains and to expose these to the
upper layers is an instrumental architectural tool, facilitating the delivery of FH and BH services. Identifying
optimal VIs in terms of both topology and resources includes:
• Ordering, referred to as SC, of the relevant functions (VNF or PNF) that need to be applied to the
traffic flows traversing the VIs.
• Estimating the virtual resources required to support SC and executing the corresponding
applications over the PI.
• Mapping of the virtual resources to the physical resources.
This process is shown at Figure 2, upper part. In this example two VIs, corresponding to different tenants,
are able to support independently FH and BH functions over a common infrastructure.
In 5G-XHaul we assume a multi-technology transport network interconnecting RUs and end-users with a
set of general-purpose servers hosted by the CU (as showcased e.g. by Alcatel-Lucent, Intel, China Mobile and
Telefónica, Mobile World Congress 2015). To support virtual FH (VFH) services, over the 5G-XHaul
infrastructure, RU demands are forwarded to a shared CU, hosting a set of sliceable and virtualised servers for
processing. Compute resources are responsible for executing the various BB functions in a predefined order
(left part of Figure 2). Based on the split option adopted, these functions can be partly executed locally at the
RU or centrally at the CU. The split choice dictates the processing allocation, to local and central compute
resources, and enforces the corresponding SC graph. A graphical representation of a typical virtual BH (VBH)
service that supports content delivery (CDN) to end-users is shown in Figure 2. The SC graph indicates that
IEEE COMMUNICATIONS MAGAZINE, IN PRESS, 2017
the VBH service allows mobile traffic, generated at the wireless access domain, to traverse a hybrid multi-hop
wireless/optical transport network, before it reaches the compute resources.
To evaluate the performance of this type of infrastructure and the proposed architecture, we have developed
a mathematical framework, based on MOP, for the integrated wireless and optical network domains,
considering the data plane described in Section II.A and the details of the compute resources required. Our
study focuses on optimal planning of VFH and VBH infrastructures, in terms of both topology and resources
considering overall power consumption and end-to-end delays. To identify the best performing FH and BH
VIs, detailed power consumption and end-to-end delay models are considered. These models describe the
details of the optical and wireless network as well as the compute domains and the associated interfaces, [9].
The results obtained, focus on the specific use case of joint FH and BH optimisation, assuming delivery of
CDN services to the end-users.
The joint VFH/VBH design problem considers also a set of constraints ensuring efficient and stable
operation of the planned Vis, summarized below:
• VFH and VBH infrastructures have specific requirements. In response to this, different VNFs are
grouped and orchestrated in the form of SCs with specific processing and network requirements. To
realize an SC, sufficient network and processing capacity must be allocated to the planned VIs for
the interconnection and deployment of VNFs. The order of VNF processing is defined by the
corresponding SC.
• Reservation of physical resources to support SC depends on the users’ mobility model assumed, the
size of the wireless cells and the traffic model adopted. Ideally, 100% overprovisioning of both
network and compute resources across neighboring cells can guarantee seamless handovers.
However, to improve resource efficiency the reservation of resources, residing in adjacent cells, can
be linked to handoff probability. The amount of resources leased in the wireless domain is assumed
to be an increasing function of the handoff probability [12]. Given that both the RUs and the end-
users need to be supported by remotely located compute resources, the additional resource
requirements also propagate in the transport network and the compute domain.
• The VI planning process considers a number of functions across different domains including flow
conservation, mapping, aggregation and deaggregation of traffic.
• Given that both FH and BH services require compute processing the associated impact in the overall
infrastructure evaluation is considered. Therefore, the traffic associated with these services is
IEEE COMMUNICATIONS MAGAZINE, IN PRESS, 2017
mapped, not only to network resources, but also to compute resource requirements. This introduces
an additional constraint, linked with the conversion of network-to-compute resource requests. To
achieve this, a mapping parameter, defined as “network-to-compute” parameter, is introduced
providing the ratio of network requirements (in Mbps) and computational requirements (in
Operations Per Second (OPS)), of a specific service demand. This parameter takes low values, for
cloud services requiring high network bandwidth and low processing capacity (e.g. video
streaming). On the other hand, it takes high values for tasks requiring intensive processing and low
network bandwidth (e.g. data mining, Internet-of-Things). Regarding BH services, the Standard
Performance Evaluation Corporation recently established the Cloud subcommittee to develop
benchmarks able to measure these parameters. Taking a similar approach the authors in [14]
measured the average ratio between computational and network bandwidth requirements for various
“Big Data” analytics workloads. Similarly, FH services require specific computing resources to
support BB processing. The processing power depends on the details of the BBU [3]-[4] including
processing tasks related to Fast Fourier Transforms (FFT), error correction, processing-resource
mapping/de-mapping etc. calculated in Giga OPS (GOPS). The resulting processing power depends
on the LTE system configuration [4].
• Our analysis takes into consideration end-to-end delays for specific services e.g. FH or real time
BH services. In highly loaded heterogeneous networks, such as the 5G-XHaul solution, end-to-end
delay can be greatly influenced by queuing delays associated with the interfaces across the
infrastructure domains. In this context, the choice of suitable queuing and scheduling policies at the
interfaces offers significant delay benefits. Traditionally, these systems can be mathematically
modeled applying queuing theory and open/closed mixed queuing networks. However, such a
model is not able to abide to the strict FH latency constraints. To address this issue, the queuing
delays for the VFH are modeled under worst case operational conditions, using network calculus
theory [13].
• Flexible processing splits refer to the choice of a single split option for every time instance. Once
the split option has been selected, the corresponding SC is applied across the network, deploying
specific network and compute resources as dictated by the relevant split. [3], [4].
Depending on the functional split, some of the processing is performed by compute resources either at a local
cloudlet [12] 𝑐, 𝑐𝜖𝒞 (𝒞 denotes the set of cloudlets) with cost 𝑤% per GOPS or at a remote regional DC 𝑠𝜖𝒮 (𝒮
denotes the set of DCs) with cost 𝑤( per GOPS. Assuming that the cost for operating FH capacity 𝑢*+,- of
IEEE COMMUNICATIONS MAGAZINE, IN PRESS, 2017
physical link 𝑒𝜖ℰ (ℰ: set of physical links) is 𝑤- and 𝜋*+,( 𝜋*+,% are the BBU processing capacities at the
remote server and the cloudlet, respectively, the optimal VFH infrastructure is determined by minimizing the
following cost:
𝑚𝑖𝑛 𝒱ℱℋ(𝒖, 𝝅) = 𝑤--=ℰ
𝑢*+,- + 𝑤(𝜋*+,((=𝒮
+ 𝑤%𝜋*+,%%=𝒞
(1)
subject to the constraints described above.
As the optical transport requires very small amount of power to operate, higher pooling gains are expected
when C-RAN solutions are adopted compared to traditional RANs. The impact of centralization is expressed
through transport network overloading introduced by FH services, leaving limited resources for BH services.
In view of this we set a secondary optimization objective with the aim to minimize BH end-to-end delay, subject
to demand processing and capacity constraints:
min𝒱ℬℋ(u, π) =1
𝒰H − uJK,H − uLK,HHMℰ+
1𝛱( − 𝜋*+,( − 𝜋O+,(PM𝒮
+1
𝛱% − 𝜋*+,% − 𝜋O+,%QM𝒞(2)
where 𝑢O+,-, 𝜋O+,( represent the BH related network and server capacity, respectively, 𝒰-, is the total capacity
of 𝑒 and 𝛱(, 𝛱% are the total processing capacity of the DC 𝑠 and the cloudlet 𝑐, respectively.
The MOP described through (1)-(2) can be written as minℱ 𝒖, 𝝅 = 𝒱ℱℋ(𝒖, 𝝅), 𝒱ℬℋ(𝒖, 𝝅) subject
to the previously discussed constraints. This problem is then transformed from an MOP into a single objective
problem using the Pascoletti-Serafini scalarization technique [10] and solved using Lagrangian Relaxation. In
the following section, the performance of the overall architecture is evaluated in terms of power consumption
and service delay adopting a realistic network topology and actual traffic statistics [15].
IV. PERFORMANCE EVALUATION
The network topology assumed is the Bristol 5G city infrastructure (Figure 3 (a)). In this infrastructure a set
of 50 APs are evenly distributed across a 10x10 km2 area. APs are backhauled through microwave point-to-
point links and TSON is adopted for the optical transport. TSON deploys a single fiber per link, 4 wavelengths
of 10Gbps each per fiber and minimum bandwidth granularity of 100Mbps. In the present study, 𝑤- is
associated with the power consumption of link 𝑒𝜖ℰ. Power consumption figures for TSON can be found in [7],
[13]. The microwave transceivers considered, have 2Gbps bandwidth and their power consumption is 45W
(Huawei OptiXRTN310).
IEEE COMMUNICATIONS MAGAZINE, IN PRESS, 2017
a) b)
c)
d)
e)
Figure 3 a) Bristol 5G city network topology with mmWave backhauling, b) Snapshot of spatial traffic load and c)
averagetraffic/BS based on the dataset[10]during 8/2012, d)-e) Total power consumption and total service delay over time for the
traditional RAN.
mmWavelinks Opticalfiber TSONnodes LargescaleDC Distance (km)
Dist
ance
(km
)
0 2.5 5 7.5 100
2.5
5
7.5
10
0
10
20
30
40
50
60
70
80
90
100
Mbps
0 10 20 30 40 50 60 70 80 90 1000
50
100
150
time (h)
Aver
age
traffi
c/ce
ll (M
bps)
IEEE COMMUNICATIONS MAGAZINE, IN PRESS, 2017
Furthermore, a 2×2 MIMO scheme with adaptive number of transmission elements, carrier frequency 2.6GHz,
20 MHz bandwidth adjustment and capacity per cell up to 201.6 Mbps has been considered. Mobile users are
distributed and generate traffic over the serviced area according to real datasets reported in [15] (Figure 3 b).
The following two scenarios are studied:
a) “Traditional RAN,” where power consumption per AP ranges between 600 and 1200 Watts under idle and
full load conditions, respectively, and commodity servers are used to support CDN services.
b) “C-RAN with virtual BBUs (vBBUs)” where commodity servers are used to support both FH (through the
creation of vBBUs [1]) and BH CDN services.
For C-RAN with vBBUs, where optimal split options are deployed, two types of servers have been considered:
a) small scale commodity servers (cloudlets) close to the APs and, b) commodity servers hosted by large scale
DCs (Figure 3 (a)) with an average cost equal to 2Watts/GOPS and 1.6Watts/GOPS, respectively. Details
regarding the numerical values used in the simulations are provided in [7]. Although both types of servers can
provide the necessary processing power for C-RAN and CDN services, large scale DCs provide superior
performance per Watt, compared to cloudlets. Figure 3 (d) shows that significant energy savings (ranging
between 60-75%) can be achieved adopting the C-RAN approach using the integrated wireless-optical
infrastructure, compared to traditional RAN. However, due to sharing of network resources between BH
services and high priority FH services, C-RAN leads in increased BH service delays that remain below 25 ms
(Figure 3 (e)). On the other hand, traditional RAN provides minimum end-to-end BH service delays, as no
sharing with FH services is required, but at the expense of increased power consumption due to the limited
BBU sharing.
The impact of mobility on the total power consumption and the optimal split option adopted is shown in Figure
4 (a) and (b), respectively. The call-to-mobility factor is defined as the ratio of the service holding time over
the cell residence time [13], with low call-to-mobility factor values indicating high degree of mobility. It is
known that high degree of mobility introduces additional resource requirements in the wireless domain. To
ensure seamless end-to-end connectivity between end-users, RUs and compute resources, these additional
resource requirements also propagate across the transport network and the compute domains. In Figure 4 (b) it
is observed that lower split options are beneficial for higher mobility, enabling a larger number of BB
processing tasks to be offloaded to remote DCs. Given that BB processing requirements increase with mobility,
a higher degree of centralization benefits the system, due to increased consolidation and improved performance
per watt that large scale remote DCs offer, compared to local cloudlets.
IEEE COMMUNICATIONS MAGAZINE, IN PRESS, 2017
Figure 4: Impact of mobility power consumption and optimal split option (load 18Mbps/cell).
Figure 5 (a) illustrates the impact of service requirements in terms of network and compute resources on the
optimal split option adopted. CDN services with high network-to-compute ratios (e.g. video analytics) require
significant network resources to operate, leading to overutilization of transport capacity. This effect is counter-
balanced by the selection of higher split options (i.e. options 3, 4) that require lower bandwidth for the
interconnection of RUs with CUs compared to the bandwidth requirements of lower split options (i.e. options
1, 2). The impact of the traffic load on the total power consumption is illustrated in Figure 5 (b). As expected,
for higher traffic load, the total power consumption increases, and a step-like increase is observed, above 45
Mbps per cell traffic load. Beyond this threshold, the preferable system split option becomes split 4 (rather than
split 1) and a large number of cloudlets per geographic region are activated to support BB processing
requirements.
Figure 5 Split option as a function of load for different Compute to network ratios.
1 2 3 4 5 6 74
4.2
4.4
4.6
4.8
5x 104
Call to mobility
Pow
er C
onsu
mpt
ion
(Wat
t)
1 2 3 4 5 6 71
2
3
4
Call to mobility
Split
optio
n
IEEE COMMUNICATIONS MAGAZINE, IN PRESS, 2017
Finally, the impact of the relative processing and transport network cost, on the optimal split option, is
illustrated in Figure 6. The relative local to remote processing cost, is defined as the ratio of the power
consumed for data processing at the local cloudlet over the power consumed for processing of the same data
remotely at large scale DCs. It is seen that increasing this ratio makes it beneficial to perform more processing
functions at large scale remote DCs. Thus a lower split option is preferable. To include the impact of the
network cost in this analysis, the end-to-end transmission cost is also plotted in Figure 6. As the transmission
cost increases (higher number of wireless hops), it is beneficial to perform more processing functions at the
local cloudlets and hence adopt a higher split option.
Figure 6 Split options for various processing and transmission costs.
V. CONCLUSIONS
This paper presents a converged optical-wireless 5G infrastructure proposed by the EU 5GPPP project 5G-
XHaul aiming to support jointly operational network and end-user services. The overarching architecture
proposed is aligned with the SDN reference architecture and the ETSI NFV standard. A novel MOP modeling
framework has been developed to evaluate the performance of the 5G-XHaul architecture taking into
consideration the joint support of FH and BH services. Our modeling results show that the proposed
architecture can offer significant benefits in terms of energy consumption but at the expense of end-user service
delays.
ACKNOWLEDGMENT
This work has been supported by the EU Horizon 2020 5GPPP project 5G-XHaul.
REFERENCES
[1] C-RAN. The Road Towards Green RAN. White Paper 3.0 (Dec. 2013), http://goo.gl/Sw6bfw
2040
600
12
34
Split 1
Split 2
Split 3
Split 4
Tranmission Cost/gbps (Watt)Cost local/remote processing
IEEE COMMUNICATIONS MAGAZINE, IN PRESS, 2017
[2] NokiaPress Release, Nokia accelerates centralized RAN deployment with expanded mobile fronthaul
solution #MWC16, Feb. 2016, [Online]: http://tinyurl.com/h5kmbvb
[3] U. Dötsch et al., “Quantitative Analysis of Split Base Station Processing and Determination of
Advantageous Architectures for LTE”, Bell Labs Tech. J, vol. 18, no.1, pp. 105-128, May 2013
[4] D. Wubben et al., “Benefits and Impact of Cloud Computing on 5G Signal Processing: Flexible
centralization through cloud-RAN," IEEE Signal Process. Mag., vol.31, no.6, pp.35-44, Nov. 2014
[5] N. Nikaein, M. K. Marina, S. Manickam, A. Dawson, R. Knopp, and C. Bonnet, “OpenAirInterface: A
Flexible Platform for 5G Research”, SIGCOMM Comput. Commun. Rev. Vol. 44, No. 5, Oct. 2014.
[6] ETSI GS NFV-SWA 001 V1.1.1 (2014-12), “Network Functions Virtualisation (NFV);Virtual Network
Functions Architecture”, 2014.
[7] H2020 Project 5G-Xhaul, Deliverable 2.2 “System Architecture Definition”, July 2016 [Online]:
http://www.5g-xhaul-project.eu/download/5G-XHaul_D_22.pdf
[8] Bo Han, V. Gopalakrishnan, J. Lusheng, L. Seungjoon, "Network function virtualization: Challenges
and opportunities for innovations," IEEE Commun. Mag., vol.53, no.2, pp.90-97, Feb. 2015
[9] B. R. Rofoee, G. Zervas, Y. Yan, M. Anastasopoulos, A. Tzanakaki, S. Peng, R. Nejabati, and D.
Simeonidou, “Hardware Virtualized Flexible Network for Wireless-DataCenter (invited)” , IEEE/OSA
J. Opt. Commun. Netw., March 2015, vol.3, pp A526 - A536
[10] A. Pascoletti, P. Serafini, “Scalarizing vector optimization problems”, J Optimization Theory Appl, vol.
42, pp. 499-524, 1984.
[11] M. Savi, M. Tornatore, G. Verticale, “Impact of processing costs on service chain placement in network
functions virtualization," in proc. of IEEE NFV-SDN, 2015 pp.191-197, 18-21 Nov. 2015
[12] J.-Y. Le Boudec, P.Thiran, “Network calculus:a theory of deterministic queuing systems for the
internet”, Springer, 2001
[13] A. Tzanakaki et al., “Virtualization of heterogeneous wireless-optical network and IT infrastructures in
support of cloud and mobile cloud services," IEEE Commun. Mag. , vol.51, no.8, pp.155-161, Aug.
2013.
[14] J. Chang, K. T. Lim, J. Byrne, L. Ramirez, and P. Ranganathan. “Workload diversity and dynamics in
big data analytics: implications to system designers,” in Proc. of ASBD 2012.
[15] X. Chen et al., “Analyzing and modeling spatio-temporal dependence of cellular traffic at city scale,”
in proc. of IEEE ICC, pp.3585-3591, 2015.